Hi! Does anyone know how to get the probability per letter when you inference a sentence in DeepSpeech? for example if I have:
“The sky is blue”
T = 0.7, D = 0.2, H = 0.1 … (the sum over the alphabet = 1)
H = 0.9 …
E = 0.8 …
…
Have you had a look at the API? And especially Metadata data structure?
Yes, I have already checked, but I am only able to derive the probability by word
The Metadata data structure should expose character-level information. Here in Python we leverage that to get back to the word-level: https://github.com/mozilla/DeepSpeech/blob/master/native_client/python/client.py#L38-L68, but the raw data you get access to either from C API or bindings should be character-level.
Currently we only expose per candidate transcript confidence scores. Per letter scores are possible to be exposed, and someone was working on that a while ago but sadly we never got a pull request.
Right, I was not looking carefully, confidence is attached to CandidateTranscript which has a set of Tokens, not a single one 
Sorry about the confusion @Piero_Volante